The Research Of Text Sentiment Analysis Based On Deep Learning

Sentiment analysis or opinion mining is an important task in the field of natural language processing.Sentiment analysis can be used to study people’s attitudes,perceptions,and emotions about specific entities.In recent years,sentiment analysis technology has received great attention in enterprises,governments,and organizations,and has also been widely used in society.However,in the research and application of sentiment analysis tasks,there are still the following problems: First,due to the complexity of sentences and the lack of feature extraction capabilities of sentiment analysis models,convolutional neural networks,long short-term memory neural networks,etc.Sentiment analysis models are still not very accurate in performing sentiment analysis tasks.Second,with the emergence of large-scale pre-training language models,the accuracy rate of sentiment analysis tasks has reached new highs,and the ”pre-training,fine-tuning” paradigm has become a popular paradigm for handling sentiment analysis tasks.Because the target of the model is inconsistent with the target in the pre-training stage and the fine-tuning stage,the model is not fully exploited in downstream tasks.Third,most deep learning models currently require high dataset size and cannot be effectively trained in less-sample scenarios.Aiming at the above problems,this thesis starts from improving the model structure and optimizing the model training process respectively,and proposes the following methods:1.Aiming at the problem that the existing convolutional neural network and long shortterm memory neural network do not work well in sentiment analysis tasks,a AttCLSTM sentiment analysis model is proposed.By combining convolutional neural network,long short-term memory neural network and attention mechanism,the model effectively extracts the features of each level of sentences.This theis uses Att-CLSTM to conduct experiments on the IMDb dataset and the SemEval 2010 Task 8 dataset,and achieved an accuracy of 94.34% and an F1-Score of 87.7% respectively.2.In view of the ”pre-training,fine-tuning” paradigm,due to the inconsistency of the target of the model during the pre-training and fine-tuning period,which makes the model insufficient ability to explore the downstream task,and most of the cur-rent models have high requirements for the scale of the dataset,which cannot be effectively trained in the less-sample scenario,a ”pre-training,prompt,matching”paradigm is proposed to train the sentiment analysis model.The ”pre-training,prompt” process in prompt learning is used to eliminate the difference between the goals of the pre-trained language model in the pre-training stage and the downstream task execution stage,and the workflow of prompt learning is optimized by combining the text matching scheme.The experimental results show that the sentiment analysis model trained by the ”pre-training,prompt,matching” paradigm achieves an average accuracy of 94.61% and 86.11% on the comment dataset and the bank customer intent recognition dataset,respectively,which is 2.25% and 4.55% higher than that of the model trained by the ”pre-training,fine-tuning” paradigm,respectively.